OpenClaw and Hermes Agent: the real setup for agents that work while you sleep

You commented NIGHTSHIFT, so here's the setup: which tool, which model, and the guardrails I didn't have time to explain in 45 seconds.

Quick honesty check before anything else: "it built something while you were away" is the claim this is actually good for. Neither tool prints you money. What they do is keep running after you close the laptop — which is a real, useful, and slightly more boring thing than the hook made it sound.

The two tools

OpenClaw and Hermes Agent are both free, open-source personal AI agent platforms. Neither is a model — they're the shell that keeps a model alive between messages, gives it memory, a schedule, and shell access, and lets you talk to it over chat apps instead of a browser tab.

• OpenClaw, from Peter Steinberger, is the more widely adopted of the two — 29 channels (Telegram, WhatsApp, Discord, Slack, Signal, iMessage, and more), works on macOS, Windows, and Linux.
• Hermes Agent, from Nous Research, is MIT-licensed, with the same multi-channel persistent-memory design and explicit support for running unattended "through the gateway" — reports, backups, briefings, no human watching.

Pick one to start. They solve the same problem; the differences are in the ecosystem and community around each, not the core mechanic.

Install

Both install with a one-liner. Read it before you run it — that's not a formality, it's the first guardrail.

OpenClaw, Windows:

powershell -c "irm https://openclaw.ai/install.ps1 | iex"

macOS/Linux installs are documented on openclaw.ai under Quick Start — the one-liner installs Node.js and everything else for you.

Hermes Agent, Windows:

irm https://hermes-agent.nousresearch.com/install.ps1 | iex

macOS/Linux instructions are on the same page, or grab the native desktop app for macOS, Windows, or Linux directly.

Which model actually does the thinking

This is the part the reel compressed into one line and it's worth a paragraph: the agent framework and the model are two different layers. OpenClaw and Hermes Agent are the persistence layer — memory, scheduling, the shell you're texting. They don't think; they keep something else alive and connected. Claude does the thinking in the demo, because that's the model both frameworks can orchestrate — OpenClaw's own site lists Claude and GPT side by side in its integrations row, and Anthropic's own Claude plans page documents connecting external tools and services through remote MCP connectors, which is the mechanism that makes this kind of hookup possible from Claude's side.

Practically: you're choosing a framework (OpenClaw or Hermes Agent) and a model behind it (Claude, GPT, or a local model, depending on what the framework supports and what you connect). Don't conflate the two when you're reading either project's docs.

The guardrails — set these first

This is the line in the video I flagged as the one thing that couldn't get cut for time, and it isn't padding here either.

Spending limits. Anything with API access and the ability to act can also spend. Set a hard usage cap at the provider level before you hand an agent a schedule, not after.

A sandboxed shell. OpenClaw's own site is explicit about this trade-off — full system access or sandboxed, your choice. Choose sandboxed until you've watched it work for a while and understand its actual failure modes, not the ones you imagined.

No blind access to production or payments. Read access to see what's there is different from write access to change it. Give an unattended agent the first, and gate the second behind a step you're actually present for.

If you want the fuller version of this idea — not tool-specific, but the general shape of briefing any agent you're going to leave running — I wrote that up separately: the night-shift brief. This page is the two specific tools; that one is the reusable template. Read both.

For a deeper, defense-in-depth take on containing agents before they touch anything that matters, see Contain AI agents before they reach production.

What this setup is not

It's not a passive-income system, and neither tool's site claims to be one. It's not a "drop it into your business" tool on day one — the honest version is: point it at one small, boring, repeated job, watch what it actually does, and widen access slowly. Both projects are under a year old. Expect rough edges.

Sources

• openclaw.ai — capabilities, install, channel list
• hermes-agent.nousresearch.com — capabilities, MIT license, install
• claude.com — plans, connectors, remote MCP

Verified live 28 August 2026. Product names are factual; no vendor partnership or endorsement is implied by either project.

Published 28 August 2026.